Pith. sign in

REVIEW 5 major objections 4 minor 33 references

Quasar photometric redshifts cannot be trusted without multi-component uncertainty models.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

For KiDS quasars, multi-component MDNs and BNNs reconstruct redshift distributions better than plain ANNs, but the best model depends on the data-quality regime and combined faint/missing-band cases degrade all models.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection The body is a useful, internally consistent benchmark of quasar photo-z uncertainty models, but the arXiv abstract contradicts its own Table 2 and promises an SOM/PIT analysis that the text doesn't deliver. the 5 major comments →

arxiv 2603.19882 v2 pith:SAZ7SYPJ submitted 2026-03-20 astro-ph.CO

Quasar photometric redshifts beyond the spectroscopic coverage: Uncertainty models and redshift distributions in the Kilo-Degree Survey DR5

Kacper Drabicki , Szymon J. Nakoneczny , Maciej Bilicki This is my paper
classification astro-ph.CO
keywords photometric redshiftsquasarsuncertainty quantificationmixture density networksBayesian neural networksredshift distributionKilo-Degree Surveymachine learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks which machine-learning architecture best estimates quasar photometric redshifts and their uncertainties, and which can reconstruct the true redshift distribution n(z) from predicted probability densities. Using KiDS DR5 photometry matched to DESI DR1 spectroscopy, the authors train artificial neural networks (ANNs), mixture density networks (MDNs), Bayesian neural networks (BNNs), and a self-organizing map, testing them on data that are faint, missing photometric bands, or both. They find that the ANN point predictions reconstruct n(z) poorly, that MDNs need at least two Gaussian components to capture multi-modal photo-z solutions, and that BNNs improve performance in the faint extrapolation regime but degrade it for bright sources. No model wins in every regime; the combination of faintness and missing bands defeats all approaches. Unsupervised clustering of the predicted PDFs identifies two degenerate redshift pairs.

Core claim

The central claim is that an accurate reconstruction of the true redshift distribution is unattainable without a proper uncertainty estimation framework, and that MDNs with at least two Gaussian components are essential to capture the intrinsic multi-modality of photo-z solutions. Concretely, on a clean random test set the ANN's distribution reconstruction error is about six times that of the best MDN; the single-Gaussian MDN is miscalibrated and produces more catastrophic outliers than the ANN. A three-component MDN is the best MDN overall, while a three-component BNN improves the negative log-likelihood by 0.11 on faint extrapolation data but is 0.07 worse for brighter objects. In the hard

What carries the argument

The load-bearing object is the Gaussian Mixture Model output layer attached to a neural network (MDN) or to a Bayesian network (BNN), which turns photo-z prediction into a full probability density function rather than a point estimate. The paper compares these PDFs through the negative log-likelihood at the true spectroscopic redshift, and reconstructs the redshift distribution by Monte Carlo sampling each predicted PDF. The BNN estimates epistemic uncertainty by averaging 100 stochastic forward passes. A separate diagnostic uses t-SNE and HDBSCAN clustering on the sampled PDF values to expose degenerate solutions.

Load-bearing premise

The DESI DR1 spectroscopic quasars matched to KiDS are treated as an unbiased ground truth for the faint and missing-band test sets; if DESI target selection, redshift failures, or the 1-arcsec cross-match correlate with r>22.9 or missing u-band, the reported model rankings and n(z) reconstructions reflect selection effects rather than model capability.

What would settle it

Take a spectroscopically complete sample of faint quasars from a deeper survey that does not depend on DESI targeting, and repeat the faint-extrapolation NLL comparison between the three-component MDN and BNN. If the BNN's 0.11 NLL advantage disappears, the reported benefit of Bayesian networks is an artifact of the DESI selection function. A second test: recompute the 'missing features' rankings after dropping objects with missing u-band; if the rankings change, the results are driven by the imputation of the most important band.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Tomographic analyses with photometric quasars should be based on predicted probability distributions, not on single point estimates, since the ANN's reconstructed n(z) shows step-like artifacts and six times the error of the best uncertainty-aware model.
  • At least two Gaussian mixture components are required in the output layer; a single Gaussian is miscalibrated and yields more catastrophic outliers than a plain ANN.
  • Bayesian neural networks with three components are the best choice for samples extending fainter than the training set (NLL gain 0.11), but they come with a 0.07 NLL penalty for well-covered bright sources.
  • Reconstruction remains feasible for fainter data or missing magnitudes separately, but fails when both occur together, so survey samples must be quality-controlled for combined incompleteness.
  • Clustering predicted photo-z PDFs can identify degenerate redshift pairs ((1.2,2.3) and (1.6,2.5)) that could be removed from cosmological inference.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because 76% of missing-band cases are missing the u-band, the paper's grouped 'missing features' test is effectively a u-band dropout test; a per-band analysis would probably show that missing u-band dominates the degradation and that other bands are nearly harmless.
  • The two degeneracies (1.2,2.3) and (1.6,2.5) are consistent with quasar spectral-feature aliasing (e.g., Ly-alpha vs CIV line confusion); adding infrared photometry or a prior on quasar luminosity could break these degeneracies and would be a testable improvement.
  • The regime-dependent best model suggests operational pipelines could adopt a two-stage design: a cheap classifier to identify data-quality regime, then a dedicated uncertainty model for that regime, rather than relying on a single network.
  • Because the study deliberately restricts to a well-matched spectroscopic sample, the reported numbers are an upper bound; applying these models to the full KiDS quasar photometric sample, where training–inference domain mismatch is real, would likely degrade all metrics and could change the ranking.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper compares ANN, MDN (with 1–5 Gaussian mixture components), and BNN (with three components) for quasar photometric redshift estimation and uncertainty quantification using KiDS DR5 photometry and DESI DR1 spectroscopy. Four test regimes are constructed from the combination of "random" vs "faint" (r > 22.9) samples and complete vs missing photometric bands. The models are compared via NLL, point-estimate metrics, and the MSE* of reconstructed n(z) histograms; unsupervised clustering (t-SNE + HDBSCAN) of predicted PDFs is used to identify degenerate solutions. The body's results indicate that at least two MDN components are needed, BNNs improve NLL in the faint/missing regimes, and n(z) reconstruction degrades substantially when missing features and magnitude extrapolation are combined. The abstract included with the arXiv listing, however, makes several claims—SOM benchmarking, PIT calibration, and the single Gaussian being "best-calibrated" in the faint-plus-missing regime—that are absent from or contradicted by the body and its tables.

Significance. If properly supported, the paper would provide a useful benchmark for uncertainty-aware quasar photo-z estimation in upcoming surveys such as Euclid and LSST. The controlled construction of four test regimes, the use of external DESI spectroscopic redshifts as ground truth, and the explicit NLL comparisons in Table 2 are genuine strengths; the body's central conclusion that multi-component MDN/BNN outputs are needed for good n(z) reconstruction is broadly consistent with the reported numbers. However, the manuscript in its current form contains internal inconsistencies between the abstract and the body, missing analyses that the abstract promises, and a row-labeling error in the n(z) results section. These issues must be resolved before the headline claims can be evaluated.

major comments (5)
  1. [Abstract vs §4.1, Table 2] The arXiv-listing abstract states that "in the hardest faint-plus-missing regime the single Gaussian becomes the best-calibrated." Table 2, column 4, reports NLL = 1.28 for MDN1, 0.74 for MDN3, and 0.69 for BNN3 in that regime: the single Gaussian has the worst, not the best, NLL. If "best-calibrated" is meant to refer to a PIT analysis, no such analysis appears in the body. The same abstract also announces a SOM forecast and a PIT metric, neither of which is found in Sections 3–5. This is not a cosmetic mismatch; it directly affects the paper's headline claim that the best model is regime-dependent.
  2. [§4.2, Fig. 2] The text and the figure caption disagree on which row corresponds to which test regime. The caption defines row 1 = random complete, row 2 = random missing, row 3 = faint complete, row 4 = faint missing; the text describes the "second row" as the extrapolation test with all magnitudes available (i.e., row 3) and then refers to "the random test (row 3)" (i.e., row 2). Since the regime-by-regime MSE* rankings are central to the conclusion, this mislabeling must be corrected. Also, "RSS values" should read "MSE* values."
  3. [§3.2] Missing features are imputed with the training-set maximum value of that feature, and this imputation is applied uniformly to all missing cases. Because u-band gaps dominate (76% of incomplete cases), the choice of imputed value can strongly influence the "missing feature" test sets, especially the faint+missing regime that the abstract highlights. The Conclusions mention that missing-feature patterns are not separated, but the imputation scheme itself is not tested or discussed as a sensitivity choice. At minimum, the paper should test an alternative imputation (median, conditional mean, or missingness indicators) or explicitly restrict its missing-band conclusions to this imputation procedure.
  4. [§4.1, Table 2; §4.2, Fig. 2] The key quantitative comparisons—e.g., BNN over MDN3 by 0.07/0.11 NLL and MSE* factors of 3 and 25—are reported without any measure of run-to-run or sampling variability. The authors train each model five times and select the best validation realization, but the spread across realizations is not reported. Several NLL differences are smaller than 0.1; without error bars or significance tests, the "regime-dependent best model" claim is not yet established. Please provide at least the standard deviation of each metric over the five training runs or a bootstrap over test objects.
  5. [§2.2 and Conclusions] The ground-truth redshift distribution is the spectroscopic distribution of the DESI DR1 subsample after the SPECTYPE='QSO', ZWARN==0 cuts and the 1-arcsec cross-match. DESI target selection and redshift success are not complete or unbiased at r>22.9 or for missing-band objects. The Conclusions appropriately note that the study is an upper bound because it does not probe training-inference domain mismatch, but the body repeatedly calls this the "true" redshift distribution. All n(z) claims should be qualified as applying to the DESI-matched spectroscopic sample; otherwise the abstract's strong statement about "fundamentally unattainable" reconstruction overreaches.
minor comments (4)
  1. [§4.1] "The BNN achieves the lowest NLL in all three tests" is imprecise: BNN is not the lowest on the random complete test (MDN3 has -0.69 vs BNN3 -0.62). It should read "in the three tests involving additional sources of uncertainty" or similar.
  2. [Title/Abstract] The arXiv title and the internal manuscript title differ, and the two abstract versions are materially different (one mentions SOM/PIT, the other does not). These should be reconciled before submission.
  3. [§3.4] Selecting the realization with the lowest validation loss from five training runs is a form of model selection; the manuscript should state explicitly whether this selection is applied independently for each test set or only on the validation set, and whether the reported NLL values are from that single realization.
  4. [§4.3] The text says clusters 8 and -1 show degeneracy, and cluster -1 "usually represents outliers." Please define the cluster numbering convention and the criterion used to label a cluster as "suspicious" before reporting the 0.5%/0.2% fractions.

Circularity Check

0 steps flagged

No significant circularity: evaluation is anchored to external DESI spec-z ground truth on held-out sets; self-citations are contextual, not load-bearing.

full rationale

The paper's central comparison is empirical rather than definitional. All models are trained on KiDS-DR5 photometry with DESI-DR1 spectroscopic redshift labels and evaluated on held-out random and faint-extrapolation test sets, including missing-band subsets. The NLL, PIT, point-estimate metrics, MSE* distribution comparisons, and clustering diagnostics are all computed against external spectroscopic redshifts, so the reported rankings are not constructed from the quantities being predicted. The only flagged problem in the manuscript is an internal consistency issue: the Abstract states that 'in the hardest faint-plus-missing regime the single Gaussian becomes the best-calibrated,' whereas Table 2 gives MDN 1 an NLL of 1.28 in that regime versus 0.69 for BNN 3, i.e., the single Gaussian has the worst NLL by that metric. This is a correctness/consistency concern, not a circularity concern, and it does not make the derivation equivalent to its inputs. The paper's own limitation that it is an 'upper bound' because it 'does not explicitly probe the additional effects of training–inference domain mismatch' further limits external generalization but is also not circular. Self-citations (e.g., Nakoneczny et al. 2021, Bilicki et al. 2021) are used for data-product context and prior KiDS photo-z work, not as the load-bearing justification for the present model ranking. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via self-citation. The derivation is therefore self-contained with respect to its external ground truth.

Axiom & Free-Parameter Ledger

13 free parameters · 7 axioms · 0 invented entities

The paper introduces no new physical entities or forces. Its central comparison rests on a chain of standard ML assumptions plus a set of hand-chosen hyperparameters. The most scientifically load-bearing assumptions are the unbiasedness of the DESI spectroscopic sample and the validity of maximum-value imputation for missing bands; both directly affect the out-of-distribution rankings and n(z) reconstruction claims.

free parameters (13)
  • ANN architecture = 11 hidden layers x 256 neurons
    Hand-chosen; no ablation reported; performance may depend on capacity.
  • MDN/BNN architecture = 12 hidden layers x 512 neurons
    Hand-chosen; no systematic architecture search.
  • Dropout rate = 0.2
    Set for MDN/BNN; not tuned per model.
  • Learning rate = 1e-4
    Fixed for all models with Adam; no schedule or tuning reported.
  • Early stopping patience = 30 epochs
    Hand-chosen; affects convergence and selected model.
  • Model selection across training runs = 5 runs, best validation loss
    Only best run used in tables; no seed-to-seed error bars, so rankings may be noisy.
  • Number of GMM components = 1-5, with k=3 recommended
    Component count varied and selected by test performance, not derived; abstract also claims 5-component best in some regimes.
  • Missing magnitude imputation = maximum feature value in training subset
    Ad hoc imputation; may distort missing-feature test subsets and favor models that handle imputation differently.
  • t-SNE perplexity / learning rate = 40 / 200
    Hand-chosen visualization parameters; clustering result may depend on them.
  • HDBSCAN min_cluster_size / min_samples = 50 / 5
    Hand-chosen; identified clusters are 0.5% and 0.2% of the sample, close to the minimum cluster size.
  • n(z) bin width and Monte Carlo draws = delta_z=0.3, 5000 draws
    Bin width choice affects MSE* ranking; chosen to focus on overall trends.
  • BNN forward passes = 100
    Number of MC passes used to estimate epistemic uncertainty; more passes would reduce noise.
  • MSE* normalization factor = 10^3
    Arbitrary scaling; the 'accurate/minor/significant/failure' thresholds are defined post hoc.
axioms (7)
  • domain assumption Spectroscopic redshifts from DESI DR1 are accurate and unbiased for matched quasars.
    Used as ground truth for NLL, MSE, and n(z). If DESI targeting or redshift failures are magnitude-dependent, rankings are biased.
  • domain assumption Cross-match within 1 arcsec correctly identifies the same quasar in KiDS and DESI.
    Section 2.2; positional matching may include chance coincidences, though likely small.
  • ad hoc to paper Imputing missing features with the training-set maximum preserves the test distribution for evaluating missing-band robustness.
    Section 3.2; the paper groups all missing-band patterns together and the imputation value is arbitrary.
  • domain assumption ANN/MDN/BNN training with NLL/MSE losses and early stopping yields converged, representative models.
    Standard practice; but no convergence diagnostics or seed variance are reported.
  • domain assumption The 'top 10% r>22.9' subset constitutes out-of-distribution extrapolation.
    Section 3.2; magnitude-only split may still have overlapping colors, and true OOD extent is not measured.
  • domain assumption Mean of the predicted PDF is the appropriate point estimator.
    Section 3.3; the paper states mean was best in MSE but does not show the median/mode comparison.
  • domain assumption Adding magnitude errors as features would not improve performance.
    Section 3.1; claimed without showing the experiments.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Quasar photometric redshifts beyond the spectroscopic coverage: Uncertainty models and redshift distributions in the Kilo-Degree Survey DR5." pith.science (2026). https://pith.science/paper/SAZ7SYPJ

@misc{pith2026260319882,
  author       = {Pith},
  title        = {Pith review of: Quasar photometric redshifts beyond the spectroscopic coverage: Uncertainty models and redshift distributions in the Kilo-Degree Survey DR5},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SAZ7SYPJ}},
  note         = {Machine review of arXiv:2603.19882}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Photometric redshifts (photo-$z$) and their distributions underpin cosmology with photometric quasars, tracers in angular clustering and cross-correlations. Progress requires trustworthy uncertainties, especially beyond the spectroscopic training set. We compare how machine-learning frameworks estimate quasar photo-$z$ uncertainties and reconstruct the redshift distribution $n(z)$ under controlled data-quality shifts. Using Kilo-Degree Survey DR5 photometry and DESI DR1 spectroscopic quasars, we train artificial neural networks (ANNs), Mixture Density Networks (MDNs) and Bayesian Neural Networks (BNNs) with Gaussian-mixture outputs, plus a self-organizing map (SOM) as a direct $n(z)$ estimator. We evaluate them on four subsets, with and without magnitude extrapolation and missing bands, through the negative log-likelihood, probability integral transform, point-estimate accuracy and the bias of binned $n(z)$ moments; degeneracies are sought by clustering the predicted PDFs. At least two mixture components are essential: a single Gaussian is miscalibrated and produces more catastrophic outliers than the ANN. No model performs best everywhere. On well-covered data the five-component MDN, three-component BNN and SOM reconstruct $n(z)$ almost perfectly; under faint extrapolation the BNN gives the best likelihoods, while in the hardest faint-plus-missing regime the single Gaussian becomes the best-calibrated. For point estimates and tomographic binning, uncertainty models outperform the ANN, while the SOM fails out-of-distribution. PDF clustering exposes distinct colour-redshift degeneracies likely to grow for fainter samples. The best model is thus regime- and application-dependent: multi-component MDNs or BNNs are needed for clean binning and are the only intrinsically calibrated choice on a well-covered golden sample, a first step towards a full comparison of quasar photo-$z$ pipelines.

Figures

Figures reproduced from arXiv: 2603.19882 by Kacper Drabicki, Maciej Bilicki, Szymon J. Nakoneczny.

Figure 1
Figure 1. Figure 1: Distributions of test sets over r magnitude. The faint￾extrapolation test set includes the top 10% of r values, starting at r > 22.9. The random test contains a randomly selected 10% of the objects with r < 22.9. Both test sets are comparable in size. The faint￾extrapolation set appears more prominent only because its objects are concentrated in a much narrower r range, whereas the random set is spread wid… view at source ↗
Figure 2
Figure 2. Figure 2: Spectroscopic (red) and photo-z (black) redshift distributions for ANN (column 1), MDN with 1 mixture component (column 2), MDN with 3 mixture components (column 3), and BNN with 3 mixture components (column 4). Histograms correspond to four subsets: random test data with complete features (row 1), random test data with at least one missing feature (row 2), faint-extrapolation test data with complete featu… view at source ↗
Figure 3
Figure 3. Figure 3: t-SNE projection of photo-z estimates in the random test data with complete features. Top-left: mean photo-z, top-right: photo-z standard deviation, bottom-left: r magnitude, bottom-right: clusters. For each object, we use an array of equally spaced 100 probability density function (PDF) values calculated from model output. impact of data incompleteness and domain shift on cosmologi￾cal inference. Furtherm… view at source ↗
Figure 4
Figure 4. Figure 4: Spectroscopic and photometric redshift distributions for all clusters. Article number, page 9 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 3 linked inside Pith

  1. [1]

    2015, TensorFlow: Large-Scale Ma- chine Learning on Heterogeneous Systems, software available from tensor- flow.org

    Abadi, M., Agarwal, A., Barham, P., et al. 2015, TensorFlow: Large-Scale Ma- chine Learning on Heterogeneous Systems, software available from tensor- flow.org

    2015

  2. [2]

    J., & Schwarz, D

    Alonso, D., Bellini, E., Hale, C., Jarvis, M. J., & Schwarz, D. J. 2021, MNRAS, 502, 876

    2021

  3. [3]

    2021, A&A, 653, A82

    Bilicki, M., Dvornik, A., Hoekstra, H., et al. 2021, A&A, 653, A82

    2021

  4. [4]

    2025, arXiv e-prints, arXiv:2512.16010

    Chen, J., Luo, Z., Fu, L., et al. 2025, arXiv e-prints, arXiv:2512.16010

    arXiv 2025

  5. [5]

    2015, keras,https://github.com/fchollet/keras da Costa-Luis, C

    Chollet, F. 2015, keras,https://github.com/fchollet/keras da Costa-Luis, C. 2019, The Journal of Open Source Software, 4, 1277 DESI Collaboration, Aghamousa, A., Aguilar, J., et al. 2016, arXiv e-prints, arXiv:1611.00036 DESI Collaboration, Karim, M. A., Adame, A. G., et al. 2025, arXiv e-prints, arXiv:2503.14745 D’Isanto, A. & Polsterer, K. L. 2018, A&A,...

    Pith/arXiv arXiv 2015

  6. [6]

    2013, The Messenger, 154, 32

    Edge, A., Sutherland, W., Kuijken, K., et al. 2013, The Messenger, 154, 32

    2013

  7. [7]

    M., Shanks, T., Metcalfe, N., et al

    Eltvedt, A. M., Shanks, T., Metcalfe, N., et al. 2024, MNRAS, 535, 2105 Euclid Collaboration, Scaramella, R., Amiaux, J., et al. 2022, A&A, 662, A112

    2024

  8. [8]

    2012, arXiv e-prints, arXiv:1208.4012

    Green, J., Schechter, P., Baltay, C., et al. 2012, arXiv e-prints, arXiv:1208.4012

    Pith/arXiv arXiv 2012

  9. [9]

    R., Millman, K

    Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Nature, 585, 357

    2020

  10. [10]

    2010, A&A, 523, A31

    Hildebrandt, H., Arnouts, S., Capak, P., et al. 2010, A&A, 523, A31

    2010

  11. [11]

    L., Wright, A

    Hildebrandt, H., van den Busch, J. L., Wright, A. H., et al. 2021, A&A, 647, A124

    2021

  12. [12]

    Hunter, J. D. 2007, Computing in Science & Engineering, 9, 90 Ivezi´c, Ž., Kahn, S. M., Tyson, J. A., et al. 2019, ApJ, 873, 111 John William, A., Bilicki, M., Hellwing, W. A., Nakoneczny, S. J., & Jalan, P. 2025a, arXiv e-prints, arXiv:2511.17311 John William, A., Jalan, P., Bilicki, M., et al. 2025b, A&A, 698, A276

    arXiv 2007

  13. [13]

    2024, ApJ, 964, 130

    Jones, E., Do, T., Boscoe, B., et al. 2024, ApJ, 964, 130

    2024

  14. [14]

    Kingma, D. P. & Ba, J. 2014, arXiv preprint arXiv:1412.6980

    Pith/arXiv arXiv 2014

  15. [15]

    2008, A&A, 482, 1053

    Kuijken, K. 2008, A&A, 482, 1053

    2008

  16. [16]

    2019, A&A, 625, A2

    Kuijken, K., Heymans, C., Dvornik, A., et al. 2019, A&A, 625, A2

    2019

  17. [17]

    2015, MNRAS, 454, 3500

    Kuijken, K., Heymans, C., Hildebrandt, H., et al. 2015, MNRAS, 454, 3500

    2015

  18. [18]

    R., Feng, H., et al

    Li, R., Napolitano, N. R., Feng, H., et al. 2022, A&A, 666, A85

    2022

  19. [19]

    2024, MNRAS, 535, 1844

    Luo, Z., Li, Y ., Lu, J., et al. 2024, MNRAS, 535, 1844

    2024

  20. [20]

    2017, Journal of Open Source Software, 2, 205

    McInnes, L., Healy, J., & Astels, S. 2017, Journal of Open Source Software, 2, 205

    2017

  21. [21]

    J., Alonso, D., Bilicki, M., et al

    Nakoneczny, S. J., Alonso, D., Bilicki, M., et al. 2024, A&A, 681, A105

    2024

  22. [22]

    J., Bilicki, M., Pollo, A., et al

    Nakoneczny, S. J., Bilicki, M., Pollo, A., et al. 2021, A&A, 649, A81

    2021

  23. [23]

    Newman, J. A. & Gruen, D. 2022, ARA&A, 60, 363

    2022

  24. [24]

    2019, A&A, 621, A26

    Pasquet, J., Bertin, E., Treyer, M., Arnouts, S., & Fouchez, D. 2019, A&A, 621, A26

    2019

  25. [25]

    2011, Journal of Machine Learning Research, 12, 2825 Pérez, F

    Pedregosa, F., Varoquaux, G., Gramfort, A., et al. 2011, Journal of Machine Learning Research, 12, 2825 Pérez, F. & Granger, B. E. 2007, Computing in Science and Engineering, 9, 21

    2011

  26. [26]

    C., Hickox, R

    Petter, G. C., Hickox, R. C., Alexander, D. M., et al. 2023, ApJ, 946, 27

    2023

  27. [27]

    2019, Nature Astronomy, 3, 212

    Salvato, M., Ilbert, O., & Hoyle, B. 2019, Nature Astronomy, 3, 212

    2019

  28. [28]

    2023, A&A, 675, A202 van der Maaten, L

    Vakili, M., Hoekstra, H., Bilicki, M., et al. 2023, A&A, 675, A202 van der Maaten, L. & Hinton, G. 2008, Journal of Machine Learning Research, 9, 2579 Van Rossum, G. & Drake, F. L. 2009, Python 3 Reference Manual (Scotts Valley, CA: CreateSpace)

    2023

  29. [29]

    E., et al

    Virtanen, P., Gommers, R., Oliphant, T. E., et al. 2020, Nature Methods, 17, 261

    2020

  30. [30]

    Waskom, M. L. 2021, Journal of Open Source Software, 6, 3021 Wes McKinney. 2010, in Proceedings of the 9th Python in Science Conference, ed. Stéfan van der Walt & Jarrod Millman, 56 – 61

    2021

  31. [31]

    H., Kuijken, K., Hildebrandt, H., et al

    Wright, A. H., Kuijken, K., Hildebrandt, H., et al. 2024, A&A, 686, A170

    2024

  32. [32]

    2024, AJ, 168, 244

    Zhang, C., Wang, W., Qu, M., Jiang, B., & Zhang, Y . 2024, AJ, 168, 244

    2024

  33. [33]

    2022, Research in Astronomy and As- trophysics, 22, 115017 Article number, page 8 Kacper Drabicki et al.: Modeling Quasar Photo-zDistribution and Uncertainty Fig

    Zhou, X., Gong, Y ., Meng, X.-M., et al. 2022, Research in Astronomy and As- trophysics, 22, 115017 Article number, page 8 Kacper Drabicki et al.: Modeling Quasar Photo-zDistribution and Uncertainty Fig. 4.Spectroscopic and photometric redshift distributions for all clusters. Article number, page 9

    2022

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.